Selecting and Generating Computational Meaning Representations for Short Texts

Language conveys meaning, so natural language processing (NLP) requires representations of meaning. This work addresses two broad questions: (1) What meaning representation should we use? and (2) How can we transform text to our chosen meaning representation? In the first part, we explore different meaning representations (MRs) of short texts, ranging from surface forms to deep-learning-based models. We show the advantages and disadvantages of a variety of MRs for summarization, paraphrase detection, and clustering. In the second part, we use SQL as a running example for an in-depth look at how we can parse text into our chosen MR. We examine the text-to-SQL problem from three perspectives—methodology, systems, and applications—and show how each contributes to a fuller understanding of the task.